A double-scale-pooling grouping high-voltage transmission corridor fire detection method

Abstract

The application discloses a kind of high-voltage transmission corridor fire detection methods of double-scale pooling grouping, it is related to the field of safe operation detection of power grid, step one, image data input and pre-processing, by the visible light camera carried by unmanned aerial vehicle UAV, the original image collected is uniformly adjusted to fixed size;Step two, backbone network feature extraction and enhancement, the image after standardization is input RCPGNet backbone network, and initial fire feature is extracted, step three, spatial detail perception and multi-scale fusion, the feature is input double-scale pooling grouping module arranged at the end of neck network, step four, feature fusion and detection, introduce variant multi-scale fusion pyramid feature fusion module in neck, step five, fire identification output, the application is adapted to complex and harsh environment in remote mountainous area of high-voltage transmission corridor, stably realizes mountain fire accurate detection and positioning, solves the early accurate monitoring and early warning pain point of transmission corridor mountain fire, provides technical support for energy cross-regional transmission safety.

Description

Technical Field

[0001] This invention mainly relates to the technical field of power grid safety operation detection, specifically a dual-scale pooling grouping method for detecting fires in high-voltage transmission corridors. Background Technology

[0002] High-voltage transmission lines are the core hubs for cross-regional energy transmission. Their corridors are mostly distributed in remote mountainous areas with complex terrain and harsh environments. Affected by extreme weather, difficulties in fire control in mountainous areas, and human disturbances, wildfires frequently occur in transmission corridors. In the past five years, the number of line tripping accidents caused by wildfires has increased by more than 32% annually nationwide. Wildfires in remote mountainous areas account for more than 75% of equipment damage. Wildfires can burn towers and conductors, causing tripping and grid disconnection accidents, resulting in economic losses and threats to ecological security. Achieving early and accurate monitoring and early warning of wildfires in transmission corridors is an urgent need to ensure the safe operation of the power grid.

[0003] With the development of deep learning and UAV remote sensing technology, vision-based fire detection is gradually replacing traditional manual inspection and single-point sensing monitoring. To optimize detection performance, existing research has improved long-distance feature capture capabilities through recurrent neural networks and global attention mechanisms, optimized the accuracy of local detail detection through feature enhancement and grid sampling, and improved the multi-scale target detection effect through adaptive feature fusion and multi-scale feature stitching, thus promoting the development of fire detection technology in power transmission corridors.

[0004] However, current detection technologies still have significant shortcomings: traditional convolutional receptive fields are limited, attention mechanisms are computationally complex, and they cannot effectively model the long-distance feature associations of flames and smoke; deep network extraction is prone to losing fine-grained features such as flame texture and smoke edges, and weak fire points and thin smoke are small, resulting in a high rate of missed detection; multi-scale fusion strategies are rigid, lack channel decoupling and dynamic weight adjustment, and are prone to feature redundancy and mutual interference between scale features, making it difficult to balance detection accuracy and real-time performance, and thus failing to meet the engineering requirements for early, accurate, and real-time monitoring of fires in high-voltage transmission corridors. Summary of the Invention

[0005] Based on this, the purpose of this invention is to provide a dual-scale pooling grouping method for detecting fires in high-voltage transmission corridors, in order to solve the technical problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A method for detecting fires in high-voltage transmission corridors using dual-scale pooling grouping includes the following steps:

[0008] Step 1: Image data input and preprocessing. The visible light camera carried by the UAV collects inspection images of the high-voltage transmission line corridor as the input source, and the collected raw images are uniformly adjusted to a fixed size of 640×640.

[0009] Step 2: Backbone network feature extraction and enhancement. The standardized image is input into the RCPGNet backbone network to extract the initial fire features. The feature map is then processed through the channel remote correlation module. Channel dimension feature recalibration and nonlinear mapping are used to capture the long-distance feature dependencies between flames and smoke.

[0010] Step 3: Spatial detail perception and multi-scale fusion. The spatial resolution perception module is used to perform multi-branch parallel processing and bidirectional spatial pooling on the features to capture fine-grained features of flame texture, smoke edges and weak fire points.

[0011] The feature input is set in the dual-scale pooling grouping module at the end of the neck network. The channel grouping strategy is used to decouple features at different semantic levels. Parallel average pooling and max pooling are used to capture the global fire distribution context and extract local salient fire details, respectively.

[0012] Step 4: Feature fusion and detection. A variant multi-scale fusion pyramid feature fusion module is introduced in the neck area, and the original CSP structure is replaced by a C2 module. Feature fusion is performed through multi-level Concat connections. A micro-detection head is added to the head to enhance low-level feature extraction.

[0013] Step 5: Fire identification output. Output the target categories and bounding boxes of flames and smoke in the image of the power transmission line corridor to achieve accurate detection and location of fires in the high-voltage power transmission line corridor.

[0014] Specifically, in this technical solution, the channel remote association module in step two processes the input feature map into two parallel branches. The upper branch performs preliminary feature compression through 1×1 convolution, reorganizes the features into channel hierarchical processing, and then generates channel attention weights through the Softmax function. The lower branch performs feature transformation through 1×1 convolution, maintaining the basic expression of the original features.

[0015] The features from the two branches are adaptively fused through multiplication. The fused features are then subjected to convolutional dimensionality reduction and compression, and convolutional dimension reconstruction, ultimately outputting a channel-optimized feature map.

[0016] Specifically, in this technical solution, the spatial resolution perception module in step three processes the input feature map through two parallel processing branches. The upper branch performs feature compression through 1×1 convolution to preserve spatial location information. The lower branch adopts a two-level serial process, first performing feature transformation through 1×1 convolution, and then performing feature aggregation in the horizontal and vertical directions through X-AvgPool and Y-AvgPool operations respectively, and extracting global context information through G-AvgPool.

[0017] The features from the two branches are processed by the Softmax function to generate spatial attention weights, and adaptive feature fusion is achieved through multiplication operations to output a spatially enhanced feature map.

[0018] Specifically, in this technical solution, the dual-scale pooling grouping module in step three processes the input features into multiple groups along the channel dimension, and each group of features is processed independently and in parallel. A dual-scale pooling strategy is adopted on the parallel branches, with the average pooling branch extracting global statistical information and the max pooling branch extracting local salient features.

[0019] The features of the two pooling branches are added together and fused to enhance each other. After convolutional integration and sigmoid activation, an adaptive weight map is generated. This weight map is multiplied with the original features to achieve dynamic recalibration of the features.

[0020] Specifically, in step four, the multi-level Concat connection of the variant multi-scale fusion pyramid feature fusion module adopts a progressive feature accumulation and integration method to enhance the model's ability to express features of irregular fire lines and large-scale diffuse smoke.

[0021] Specifically, in step four of this technical solution, the micro-detection head focuses on extracting low-level detailed features to accurately identify weak fire points and small-scale fire targets with thin smoke in high-voltage transmission corridor scenarios.

[0022] Specifically, the channel remote association module, through a nonlinear mapping design of feature compression and dimension reconstruction, completes the integrated output of global channel features while reducing computational overhead.

[0023] Specifically, in this technical solution, the spatial resolution perception module dynamically weakens the background noise interference of vegetation shadows and terrain textures through adaptive allocation of spatial attention weights.

[0024] Specifically, in this technical solution, the dual-scale pooling grouping module performs spatial dimension normalization on the importance coefficients and introduces scaling and offset parameters to eliminate coefficient deviations between different samples.

[0025] Specifically, in this technical solution, the channel remote association module and the spatial resolution perception module are embedded in series at the output end of the backbone network and each layer of the neck network.

[0026] In summary, the present invention has the following main beneficial effects:

[0027] It is suitable for application scenarios in remote mountainous areas, complex terrain, and harsh environments of high-voltage transmission corridors. It can effectively resist the influence of complex conditions such as extreme weather and background interference, and stably achieve accurate detection and location of wildfires. It can significantly reduce the risk of accidents such as tower burning, line tripping, and grid disconnection caused by wildfires, reduce economic losses and ecological security threats, solve the engineering pain point of early accurate monitoring and early warning of wildfires in transmission corridors, and provide reliable technical support for the safety of cross-regional energy transmission.

[0028] The channel remote association module employs a dual-branch parallel architecture for channel feature recalibration and nonlinear mapping. It accurately models channel-level long-distance feature dependencies with low computational complexity, efficiently captures global feature associations of cross-regional fire points and large-scale smoke, strengthens feature interaction between channels, breaks through the bottleneck of weak global perception and missed detection of long-distance fires in traditional methods, and improves the accuracy of fire identification and situation assessment in complex mountainous areas.

[0029] The spatial resolution perception module employs multi-branch decoupling and bidirectional spatial pooling to accurately preserve and enhance local fire features. Combined with spatial attention weight allocation, it suppresses background noise interference and adds a micro-detection head to the detection head to enhance the extraction of low-level detailed features. The entire process improves the perception capability of small-scale targets, reduces the early fire false alarm rate, and enables early identification and early warning of wildfires.

[0030] The dual-scale pooling grouping module achieves complementary fusion of global fire distribution and local fire point details. Combined with adaptive weight recalibration, it dynamically suppresses feature redundancy and enhances effective fire features. The neck uses an efficient C2 module to replace the CSP structure, and achieves progressive feature fusion through multi-level Concat. It can cover full-scale detection from pinpoint fire points to large-area fires, while reducing computational overhead, ensuring real-time inference, and solving the problems of scale feature conflict and poor fusion quality. It balances detection accuracy and real-time response capability. Attached Figure Description

[0031] Figure 1 This is a flowchart of the method steps of the present invention;

[0032] Figure 2 This is an overall framework diagram of the detection method of the present invention;

[0033] Figure 3 This is a structural diagram of the channel remote association module of the present invention;

[0034] Figure 4 This is a structural diagram of the spatial resolution sensing module of the present invention;

[0035] Figure 5 This is a structural diagram of the dual-scale pooling grouping module of the present invention. Detailed Implementation

[0036] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0037] The embodiments of the present invention will now be described.

[0038] Based on the RCPGNet network architecture, the system consists of an image preprocessing module, a backbone feature extraction network, a channel remote correlation module (CRCM), a spatial resolution sensing module (SRCM), a dual-scale pooling grouping module (DSPG), a variant multi-scale fusion pyramid neck network, and a detection head with a micro-detection head. The CRCM and SRCM are embedded in series at the output of the backbone network and in each layer of the neck network, while the DSPG is independently set at the end of the neck network. Through the collaboration of multiple modules, the system can accurately detect and locate flame and smoke targets in high-voltage transmission corridors.

[0039] In this embodiment, please refer to Figure 1 - Figure 5 As shown,

[0040] A method for detecting fires in high-voltage transmission corridors using dual-scale pooling grouping includes the following steps:

[0041] Step 1: Image data input and preprocessing. The visible light camera carried by the UAV collects inspection images of the high-voltage transmission line corridor as the input source. The collected raw images are uniformly adjusted to a fixed size of 640×640 to complete the image standardization preprocessing and ensure the consistency of data format in subsequent network processing.

[0042] Step 2: Backbone network feature extraction and enhancement. Input the standardized image into the RCPGNet backbone network to extract the initial fire features.

[0043] The channel remote correlation module and the spatial resolution perception module are embedded in series at the output end of the backbone network and each layer of the neck network. The feature map is processed through the channel remote correlation module, and the long-distance feature dependency relationship between flame and smoke is captured by channel dimension feature recalibration and nonlinear mapping.

[0044] The specific processing steps of the channel remote association module are as follows: the input feature map is divided into two parallel branches. The upper branch performs preliminary feature compression through 1×1 convolution and then reorganizes the features into channel hierarchical processing. Subsequently, channel attention weights are generated through the Softmax function. The lower branch performs feature transformation through 1×1 convolution while maintaining the basic expression of the original features. The features of the two branches are adaptively fused through multiplication. The fused features are then subjected to convolutional dimensionality reduction compression and convolutional dimension reconstruction to finally output the channel-optimized feature map.

[0045] The channel remote association module achieves the integrated output of global channel features while reducing computational overhead through a nonlinear mapping design of feature compression and dimension reconstruction.

[0046] Step 3: Spatial detail perception and multi-scale fusion. The spatial resolution perception module is used to perform multi-branch parallel processing and bidirectional spatial pooling on the features to capture fine-grained features of flame texture, smoke edges and weak fire points.

[0047] The spatial resolution perception module processes the input feature map through two parallel branches. The upper branch performs feature compression using a 1×1 convolution to preserve spatial location information. The lower branch employs a two-stage concatenated process: first, it transforms the features using a 1×1 convolution, then performs feature aggregation in the horizontal and vertical directions using X-AvgPool and Y-AvgPool operations respectively, and extracts global context information using G-AvgPool. The features from both branches are processed by the Softmax function to generate spatial attention weights, and adaptive feature fusion is achieved through multiplication operations, outputting a spatially enhanced feature map. The spatial resolution perception module dynamically weakens background noise interference from vegetation shadows and terrain textures through adaptive allocation of spatial attention weights.

[0048] The processed features are input into the dual-scale pooling grouping module at the end of the neck network. The channel grouping strategy is used to decouple features at different semantic levels. Parallel average pooling and max pooling are used to capture the global fire distribution context and extract local salient fire details, respectively.

[0049] The specific processing procedure of the dual-scale pooling grouping module is as follows: the input features are divided into multiple groups along the channel dimension, and the features of each group are processed independently and in parallel. A dual-scale pooling strategy is adopted on the parallel branches. The average pooling branch extracts global statistical information, and the max pooling branch extracts local salient features. The features of the two pooling branches are added and fused to complement each other for enhancement. After convolution integration and sigmoid activation, an adaptive weight map is generated. This weight map is multiplied with the original features to achieve dynamic recalibration of the features. The dual-scale pooling grouping module performs spatial dimension normalization on the importance coefficients and introduces scaling and offset parameters to eliminate coefficient deviations between different samples.

[0050] Step 4: Feature fusion and detection. A variant multi-scale fusion pyramid feature fusion module is introduced in the neck area. The original CSP structure is replaced by the C2 module. Feature fusion is performed through multi-level concat connections. The multi-level concat connections adopt a progressive feature accumulation and integration method to enhance the model's feature expression ability for irregular fire lines and large-scale smoke diffusion.

[0051] A micro-detection head is added to the network detection head to enhance low-level feature extraction. The micro-detection head focuses on extracting low-level detailed features to accurately identify small-scale fire targets such as weak fire points and thin smoke in high-voltage transmission corridor scenarios.

[0052] Step 5: Fire identification output. Output the target categories and bounding boxes of flames and smoke in the image of the power transmission line corridor to achieve accurate detection and location of fires in the high-voltage power transmission line corridor.

[0053] The overall network framework consists of a backbone network responsible for extracting feature information, a variant multi-scale fusion pyramid feature fusion module designed to address the challenges posed by complex forest backgrounds, and a detection layer specifically configured for the accurate detection of small-scale flames and smoke.

[0054] The channel remote association module is a novel global dependency modeling module designed to accurately model long-distance feature associations between channels and enhance the global channel feature interaction capabilities. This enables precise matching and collaborative representation of dispersed fire points and large-scale smoke features across regions. CRCM first adopts a dual-branch parallel processing architecture, decoupling the input features into an attention generation branch and an original feature preservation branch. The attention generation branch maps spatial dimension information to the channel dimension through channel flattening and feature reorganization operations, and generates channel attention weights after Softmax normalization, accurately modeling long-distance feature dependencies between channels. The original feature preservation branch completes feature transformation through 1×1 convolution, maintaining the basic expression of the original fire features. Subsequently, CRCM performs adaptive feature fusion operations, multiplying the generated channel attention weights element-wise with the transformed original features to dynamically enhance effective channel features and suppress interference from invalid background channels. Finally, the module completes the integrated output of global channel features while reducing computational overhead through feature compression and dimension reconstruction operations, significantly enhancing the model's perception accuracy of long-distance fire features and its adaptability to wide-area complex fire scenarios.

[0055] To address the unique characteristics of feature representation in fire detection tasks, a remote channel association module is proposed, such as... Figure 3 As shown, the module first divides the input feature map into two parallel branches: the upper branch performs preliminary feature compression through 1×1 convolution, then reorganizes the features into H×W×1 dimensions for channel hierarchical processing, and then generates channel attention weights through the Softmax function, effectively capturing the correlation between channels; the lower branch also performs feature transformation through 1×1 convolution, maintaining the basic expression of the original features. The features of the two branches are adaptively fused through multiplication. This fusion mechanism allows the module to dynamically adjust the importance of features according to the channel attention weights. The fused features are first compressed by convolution to reduce the amount of computation while extracting key features, and then the feature dimensions are reconstructed through convolution, finally outputting the channel-optimized feature map X'.

[0056] The entire module, through the synergistic effect of two branches, not only models the long-range dependencies between channels but also maintains the expressive power of local features. Meanwhile, the cascaded design of multi-level convolutions effectively controls computational overhead while ensuring feature extraction quality. This not only enhances the model's ability to perceive subtle features in fire scenes but also highlights the feature representation of key channels through an attention mechanism, ultimately forming a feature optimization module that combines efficiency and performance. The specific process is as follows:

[0057]

[0058] In this formula, From the input channel dimension Compress to Generate value features that contain the channel content; Reduce the number of channels of the input features from Reduced to Generate the query features required for attention calculation; Change channel weights from Restore to the original Channels enable the reconstruction of channel dimensions.

[0059] Subsequently, the value features and query features can be generated using the following formula:

[0060]

[0061]

[0062] Value features are used to characterize the core features of the channel, while query features are used to generate attention weights; the attention weights are normalized according to the following:

[0063]

[0064] Finally, by applying the attention weights to the value features through weighted summation, the channel-optimized features can be obtained:

[0065]

[0066] Ultimately, the global context features are aggregated into the features of each query location through element-wise multiplication.

[0067] To address the need for spatial feature representation in fire detection tasks along high-voltage transmission line corridors, a Spatial Resolution Sensing Module (SRCM) is proposed, such as... Figure 4 As shown, this module enhances the model's ability to perceive spatial relationships through multi-path feature processing and spatial attention mechanisms;

[0068] Specifically, the input feature map X first goes through two parallel processing branches: the upper branch performs feature compression through 1×1 convolution, followed by C / 2×H×W feature transformation to preserve spatial location information; the lower branch adopts a two-stage cascaded processing flow, first performing feature transformation through 1×1 convolution, then performing X-AvgPool and Y-AvgPool operations to aggregate features in the horizontal and vertical directions respectively, and extracting global context information through G-AvgPool. This multi-scale pooling strategy effectively captures spatial dependencies in different directions.

[0069] The features from the two branches are processed by the Softmax function to generate spatial attention weights, and adaptive feature fusion is achieved through multiplication. The fused features are further processed by 1×1 convolution to adjust the channel dimensions, and finally output a spatially enhanced feature map X'. Through the extraction and fusion of multi-scale spatial information, not only is the model's ability to understand spatial relationships in fire scenes enhanced, but the attention mechanism also highlights the feature representation of key regions, ultimately forming an optimization module that can effectively improve the spatial feature representation capability. The specific process is as follows:

[0070]

[0071] in, It is a value feature transformation matrix used to compress the channel dimension.

[0072] It is a query feature transformation matrix used to generate attention query features. It is a spatial weight recovery matrix used to restore the compressed dimension to the original number of channels; Indicates global average pooling. This indicates pooling along the x-direction. This indicates pooling along the y-direction. This represents the Sigmoid activation function.

[0073] For input features The definition of global pooling is:

[0074]

[0075] Where H and W represent the height and width of the feature map, respectively; the symbols b, c, i, and j represent the batch size, number of channels, height index, and width index, respectively. Pooling along the x-direction is defined as:

[0076]

[0077] Pooling along the y-direction is defined as:

[0078]

[0079] The final features are output by flattening the fused features into vectors and applying Softmax to calculate the weights, then weighting the value features, and finally applying the Sigmoid activation function to the weighted features.

[0080] To address the dynamic scale variations of flames and smoke in high-voltage transmission line corridor fire scenarios, a dual-scale pooling grouping module (DSPG) is proposed, such as... Figure 5 As shown, this module first decouples the input feature X through channel grouping. This grouping strategy allows different channel groups to focus on feature expressions at different semantic levels. On the main path, the feature first undergoes channel grouping, and each group of features is processed independently to maintain its specificity. The parallel branch adopts a dual-scale pooling strategy: the average pooling branch is responsible for extracting global statistical information and capturing large-scale spatial context; the max pooling branch focuses on extracting local salient features and effectively identifying small-scale fire points and initial smoke. The features of these two pooling branches are enhanced by additive fusion. Then, the features are integrated through 2×2 convolution and enhanced by position-by-position product operation. After standardization and sigmoid activation, an adaptive weight map is generated. This weight map achieves dynamic recalibration of features through multiplication with the original features. It can not only effectively handle the scale changes from initial fire points to large-area fire spread, but also accurately capture dynamic features from local smoke to large-scale spread.

[0081] The core design of the Dual-Scale Pooling Grouping (DSPG) module lies in dividing the input features into multiple groups along the channel dimension. Each group of features is processed independently and in parallel to capture fine-grained semantic information. It is divided into G groups along the channel dimension. The dimensions after grouping are, where, Represents grouping characteristics;

[0082] Taking the first group as an example, it will enter a two-branch pooling process. The upper branch generates a global statistical description of the group's features through global average pooling. The specific process is as follows:

[0083]

[0084] The next branch combines convolution with double pooling to extract local saliency and statistical information from the feature set. These features are then fused to generate semantic features. Based on these semantic features, the model generates importance coefficients for each feature, obtained through dot product operations. This calculation process effectively quantifies global semantic features. With local features Similarity between them:

[0085]

[0086] To avoid bias in the coefficients across different samples, this section normalizes the importance coefficients in the spatial dimension:

[0087]

[0088] To ensure that the normalization operation of the insertion has the same effect as the identity transformation, a set of parameters is introduced for each coefficient. and This is used to scale and offset the normalized values.

[0089]

[0090] In a single DSPG unit, and The number of parameters is equal to the number of groups G, usually in the tens (generally 32 or 64), which is negligible compared to the millions of parameters in the entire network. Finally, to obtain the enhanced feature vector, the original features... The importance coefficients generated by the Sigmoid function are scaled in the spatial dimension:

[0091]

[0092] Furthermore, all the enhanced features together constitute the final feature set. .

[0093] When processing complex remote sensing images of fires along high-voltage transmission line corridors, the design of DSPG has advantages over traditional convolutional neural networks. Traditional convolutional neural network methods focus on using aggregated hierarchical semantic sub-features to complete the feature representation of complex targets. Sub-features are often distributed in the feature vectors of each layer in the form of groups, corresponding to different semantic entities. However, activating these sub-features is often interfered with by similar patterns and noisy backgrounds in space, resulting in target localization and recognition errors.

[0094] The proposed DSPG addresses this problem by generating an attention factor at each spatial location, which adjusts the importance of each independent sub-feature, enabling each feature group to autonomously enhance its learned multi-scale flame and smoke feature representations while suppressing potential noise interference.

[0095] Although embodiments of the present invention have been shown and described, these specific embodiments are merely explanations of the invention and are not intended to limit it. The specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. After reading this specification, those skilled in the art may make modifications, substitutions, and variations to the embodiments as needed without departing from the principles and spirit of the invention, but such modifications, substitutions, and variations are protected by patent law as long as they are within the scope of the claims of the present invention.

Claims

1. A method for detecting fires in high-voltage transmission corridors using dual-scale pooling grouping, characterized in that, Includes the following steps: Step 1: Image data input and preprocessing. The visible light camera carried by the UAV collects inspection images of the high-voltage transmission line corridor as the input source, and the collected raw images are uniformly adjusted to a fixed size of 640×640. Step 2: Backbone network feature extraction and enhancement. The standardized image is input into the RCPGNet backbone network to extract the initial fire features. The feature map is then processed through the channel remote correlation module. Channel dimension feature recalibration and nonlinear mapping are used to capture the long-distance feature dependencies between flames and smoke. Step 3: Spatial detail perception and multi-scale fusion. The spatial resolution perception module is used to perform multi-branch parallel processing and bidirectional spatial pooling on the features to capture fine-grained features of flame texture, smoke edges and weak fire points. The feature input is set in the dual-scale pooling grouping module at the end of the neck network. The channel grouping strategy is used to decouple features at different semantic levels. Parallel average pooling and max pooling are used to capture the global fire distribution context and extract local salient fire details, respectively. Step 4: Feature fusion and detection. A variant multi-scale fusion pyramid feature fusion module is introduced in the neck area, and the original CSP structure is replaced by a C2 module. Feature fusion is performed through multi-level Concat connections. A micro-detection head is added to the head to enhance low-level feature extraction. Step 5: Fire identification output. Output the target categories and bounding boxes of flames and smoke in the image of the power transmission line corridor to achieve accurate detection and location of fires in the high-voltage power transmission line corridor.

2. The method for detecting fires in high-voltage transmission corridors using dual-scale pooling grouping according to claim 1, characterized in that, The specific processing procedure of the channel remote association module in step two is as follows: the input feature map is divided into two parallel branches. The upper branch performs preliminary feature compression through 1×1 convolution and then reorganizes the features into channel hierarchical processing. Subsequently, channel attention weights are generated through the Softmax function. The lower branch performs feature transformation through 1×1 convolution to maintain the basic expression of the original features. The features from the two branches are adaptively fused through multiplication. The fused features are then subjected to convolutional dimensionality reduction and compression, and convolutional dimension reconstruction, ultimately outputting a channel-optimized feature map.

3. The method for detecting fires in high-voltage transmission corridors using dual-scale pooling grouping according to claim 1, characterized in that, The specific processing procedure of the spatial resolution perception module in step three is as follows: the input feature map goes through two parallel processing branches. The upper branch performs feature compression through 1×1 convolution to retain spatial location information. The lower branch adopts a two-level serial process. First, the feature is transformed through 1×1 convolution. Then, the X-AvgPool and Y-AvgPool operations are used to aggregate features in the horizontal and vertical directions, respectively. Finally, the global context information is extracted through G-AvgPool. The features from the two branches are processed by the Softmax function to generate spatial attention weights, and adaptive feature fusion is achieved through multiplication operations to output a spatially enhanced feature map.

4. The method for detecting fires in high-voltage transmission corridors using dual-scale pooling grouping according to claim 1, characterized in that, The specific processing procedure of the dual-scale pooling grouping module in step three is as follows: the input features are divided into multiple groups along the channel dimension, and each group of features is processed independently and in parallel. A dual-scale pooling strategy is adopted on the parallel branch, with the average pooling branch extracting global statistical information and the max pooling branch extracting local salient features. The features of the two pooling branches are added together and fused to enhance each other. After convolutional integration and sigmoid activation, an adaptive weight map is generated. This weight map is multiplied with the original features to achieve dynamic recalibration of the features.

5. The method for detecting fires in high-voltage transmission corridors using dual-scale pooling grouping according to claim 1, characterized in that, In step four, the multi-level Concat connection of the variant multi-scale fusion pyramid feature fusion module adopts a progressive feature accumulation and integration method to enhance the model's ability to express features of irregular fire lines and large-scale smoke diffusion.

6. The method for detecting fires in high-voltage transmission corridors using dual-scale pooling grouping according to claim 1, characterized in that, The micro-detection head in step four focuses on extracting low-level detailed features to accurately identify weak fire points and small-scale fire targets with thin smoke in high-voltage transmission corridor scenarios.

7. The method for detecting fires in high-voltage transmission corridors using dual-scale pooling grouping according to claim 2, characterized in that, The channel remote association module achieves the integrated output of global channel features while reducing computational overhead through a nonlinear mapping design of feature compression and dimension reconstruction.

8. The method for detecting fires in high-voltage transmission corridors using dual-scale pooling grouping according to claim 3, characterized in that, The spatial resolution perception module dynamically weakens background noise interference from vegetation shadows and terrain textures by adaptively allocating spatial attention weights.

9. The method for detecting fires in high-voltage transmission corridors using dual-scale pooling grouping according to claim 4, characterized in that, The dual-scale pooling grouping module performs spatial dimension normalization on the importance coefficients and introduces scaling and offset parameters to eliminate coefficient bias between different samples.

10. The method for detecting fires in high-voltage transmission corridors using dual-scale pooling grouping according to claim 1, characterized in that, The channel remote association module and the spatial resolution perception module are connected in series and embedded in the backbone network output and each layer of the neck network.

Images